search

gorkemcanates / Dual-Cross-Attention

Official Pytorch implementation of Dual Cross-Attention for Medical Image Segmentation
MIT License
102 stars 12 forks source link

dca_utils.py:PatchEmbedding #4

Closed wyb1022 closed 1 year ago

wyb1022 commented 1 year ago

self.projection = nn.Conv2d(in_channels=in_features, out_channels=out_features, kernel_size=size // patch_size, stride=size // patch_size, padding=(0, 0), ) Hi, Is this really patch_size here? I don't see patch_size being passed as a parameter to PatchEmbedding.

gorkemcanates commented 1 year ago

patch_size parameter here is actually the number of patches. It looks like there has been a mistake with the naming, thanks for noticing. Since we aim at obtaining a fixed number of patches for different encoder stages, it was easier to pass size of the i'th encoder and number of patches to calculate patch_size (patch size = size//patch_size in our code) for each i'th encoder features. Also note that our model does not use convolution as for the patch embedding it however uses average pooling which can be found as PoolEmbedding class under dca_utils script.

wyb1022 commented 1 year ago

thanks so much